-
Code for the article published in Neurocomputing
-
Change in the last name of the first author [Nascimento -> Castro]. I usually use my middle name on Github, Linkedin and other platforms, so it seemed right to use it also in the paper.
-
Code refactoring
-
Renaming folder and files :[experiments] -> [notebooks]; [baseline.py] -> [arima.py]; [dataset-variables.py] -> [settings.py]
-
Architecture nomenclature: we have rephrased the nomenclature of your proposed architecture for clarity, since your approach is not similar to the encoder-decoder architecture for 3D CNN.
-
-
Major changes (introduce new features and change the old version in incompatible ways)
-
We introduce a new method in our model that prevents it from violating the temporal order (STConvS2S-R). We continue using in another model (STConvS2S-C) the causal convolution (commonly used in 1D CNN), where we adapt it in spatiotemporal problems using 3D CNN.
-
We devise a temporal generator block that presents a new use of transposed convolutional layers to generate an output sequence whose length may be longer than the length of the input sequence.
-
-
Minor changes
- We include two more comparisons against our proposed methods (PredRNN and MIM) and a series of ablation studies.